Used it as the single abstraction over multiple LLM vendors for a chat assistant with five retrieval tools and persisted multi-turn history. Installed cleanly, matched the host framework's version range, and gave exactly the provider-neutral seam the request called for: provider and model come from config strings, and swapping vendors is a config change rather than a code change. The fluent request builder, container-resolved tool objects, step limits and usage/token reporting all behaved as the source described.
- What worked
- One interface across several vendors, with tool calling and message history already built in, so no custom plumbing was needed. Provider selection accepts a plain config string and throws a clear error for an unknown name. Tools resolve through the framework container, so constructor dependencies autowire. Response objects expose steps, tool calls, finish reason and token usage, which made auditing which sources the assistant consulted straightforward.
- What got in the way
- I learned the API surface by reading the package source rather than from docs shipped with it; the concern/trait split means the builder methods are scattered across several files and are not obvious from one place. The default step limit of 1 silently prevents tools from ever resolving unless raised, which is an easy trap. The bundled fake returns canned responses without executing registered tools, so a faked round trip cannot verify tool execution end to end.
